Columbia Student Wins National Award for Novel Use of Deep Learning

Chia-Hao Liu, a doctoral candidate in Applied Physics at Columbia, won the Margaret C. Etter Student Lecturer Award from the American Crystallographic Association during its recent 2019 annual meeting.

Liu was recognized for using machine learning techniques, especially deep learning, to study nanomaterials, which are being widely used now in everyday products. And his research can help improve the performance of products such as solar cells and batteries that would benefit from improved nanomaterials. As a graduate research assistant working for Professor Simon Billinge, a DSI member, Liu analyzes complex material structures by combining the techniques of x-ray and neutron diffraction with artificial intelligence. His research is leading to a better understanding of the materials on the nanoscale.

The award was established to honor the memory of Margaret C. Etter, a major contributor to the field of organic solid-state chemistry. Every year, the officers from each Scientific Interest Group (SIG) of the ACA invite one student to receive the award and to present a lecture in one of the sessions organized by that SIG.

Liu’s lecture, “Exploring the symmetry encoding in the atomic pair distribution function with convolutional neural networks,” detailed his research on applying machine learning techniques to study structures at the nanoscale.  His work is also described in a paper that is highlighted on the cover of the July 2019 issue of the journal Acta Crystallographica Section A: Foundations and Advances.

“An in-depth understanding of the material structure is the key for discovering new materials that can bring huge impacts on our lives,” says Liu. “And since nowadays there’s a great demand for smaller cell phones with longer battery life, or solar cell panels with higher conversion efficiencies, there's a corresponding demand for understanding the nanostructures that are related to the performance of these devices.”

To study the nanostructures with deep learning technique, which requires an enormous amount of data and heavy computations, Liu uses “Habanero,” the latest high-performance computing clusters at Columbia University. He computed the atomic pair distribution function – a 1D representation of the 3D material structure – from hundreds of thousands of known materials, and trained the convolutional neural network (CNN) to predict the symmetry, namely letting the CNN model solve the puzzle. He’s partnering with Daniel Hsu, a professor of computer science and Qiang Du, a professor of applied mathematics, both members of DSI, who are helping to develop the machine learning models. The use of machine learning in physical science is taking off, says Liu, with scientists using ML models in an attempt to solve an array of problems. And he’s happy about his group’s success in applying a CNN to understand nanostructures.

“The preliminary success of our deep learning model shows not only the potential of characterizing materials with nanostructures but also the great potential of applying machine learning techniques to solve challenging problems in physical science in general.”

—By Robert Florida

550 West 120th Street, Northwest Corner Building, Suite 1401, New York, N.Y. 10027    212.854.5660
©2020 Columbia University